Client-specific anomaly detection for face presentation attack detection

摘要: Abstract One-class anomaly detection approaches are particularly appealing for use in face presentation attack (PAD), especially an unseen scenario, where the system is exposed to novel types of attacks. This work builds upon anomaly-based formulation problem and analyses merits deploying client-specific information spoofing detection. We propose training one-class classifiers (both generative discriminative) using representations obtained from pre-trained deep Convolutional Neural Networks (CNN). In order incorporate information, a distinct threshold set each client based on subject-specific score distributions, which then used decision making at test time. Through extensive experiments different systems, it shown that model construction as well boundary selection) improves performance significantly. also show solutions have capacity perform or better than two-class scenarios. Moreover, CNN features models trained recognition appear discard discriminative traits less capable PAD compared CNNs generic object task.